PoUMind โ Model Weights
Production model weights for PoUMind, a multimodal emotion analysis + HTP drawing analysis platform for adolescent mental health.
- ๐ฆ Code & full documentation: https://github.com/yjkim7825/PoUMind
- ๐ POSCO Youth AIยทBigData Academy 31st, Class C Team 1 (2025.11โ12)
โ ๏ธ Disclaimer: Research / educational demo only. Not a substitute for professional psychological counseling or diagnosis. Outputs may be inaccurate. In a crisis, contact a professional helpline (in Korea: 109 / 1388 / 1577-0199; elsewhere: https://findahelpline.com).
Models
| File | Size | Architecture | Purpose |
|---|---|---|---|
checkpoints/resnet34_fold5_best.pth |
81 MB | ResNet34 | Facial emotion (image branch, deployed) |
checkpoints/best_resnet18_so.pth |
44 MB | ResNet18 | Speech emotion (Mel-Spectrogram) |
checkpoints/fusion_best_model.pth |
78 MB | ResMLP | Multimodal late fusion (image + audio) |
checkpoints/best.pt |
6 MB | YOLO | HTP object detection (house/tree/person) |
htp/yolov8m.pt |
50 MB | YOLOv8m | HTP detection base model |
6 emotion classes: Angry, Anxious, Joy, Neutral, Sad, Surprise
Reported performance
- Facial (ResNet50, training): Val Acc 0.9294 / AUROC 0.9947
- Speech (ResNet18 + Mel): Val Acc 0.7460
- Multimodal fusion (ResMLP 6656): Test Acc 0.89 / F1 0.87 (+13%p over single modality)
- HTP detection (interim, YOLOv8): mAP@50 0.957
Usage
from huggingface_hub import hf_hub_download, snapshot_download
# Download a single weight
path = hf_hub_download(
repo_id="yjkim7825/poumind-models",
filename="checkpoints/fusion_best_model.pth",
)
# Or download all weights
local_dir = snapshot_download(repo_id="yjkim7825/poumind-models")
Place the downloaded checkpoints/ files under web/data/checkpoints/, best.pt under web/data/, and yolov8m.pt under HTP/models/ to run the PoUMind server (see the GitHub README).
License
MIT License โ see the project repository.